How it works
The whole machine, end to end.
How each day's literature is fetched, de-duplicated and ranked; what every signal on a card means and where it comes from; exactly which parts use AI and which don't; and, plainly, how it can be wrong. The ranking is auditable, the signals come straight from PubMed, and none of it is a black box.
Section 1
The daily pipeline
Six stages run every day, unattended. Stages 1–3 and 5 involve no AI and no API key of any kind; only the embedding step calls a paid model, and it is a fraction of a cent per article, once.
-
Fetch
For each specialty, curated PubMedThe free U.S. National Library of Medicine database of biomedical literature, over 30 million article citations. queries (through the NCBI E-utilitiesThe NCBI's public programming interface (API) for searching and downloading PubMed records automatically, no manual searching.) and journal table-of-contents sweeps pull the newest articles, weighted toward the strongest evidence types.
-
De-duplicate & index
New records are matched against what's already stored (identifier first, then title and journal) and written to a single SQLiteA small, fast database that lives in a single file on disk, with no separate database server to run or maintain. file, with an FTS5SQLite's built-in full-text search engine. It indexes every word in the titles and abstracts so keyword search returns instantly. full-text index built over titles and abstracts so keyword search returns instantly.
-
Rank
Each article is scored for every specialty it could belong to, and the scores land in a unified index, so "top nephrology this week" is an indexed lookup rather than a scan of the corpus.
-
Embed
Every article is turned into an embeddingA list of numbers that represents a text's meaning, so two papers on the same topic land close together in that mathematical space. It powers search-by-meaning and related papers. (VoyageVoyage AI's embedding model. It reads each abstract and turns it into a vector of numbers that captures its meaning.,
voyage-4-large, 1024 dimensions). That vector powers semantic search, related papers, duplicate detection and club-tuned ranking. -
Enrich
Out of band, citation impact, retraction status, open-access full text and funding details are backfilled from public sources, so a card gets richer over the days after it first appears.
-
Serve
Your feed shows only what clears the relevance bar you set, flags what's new since your last visit, and is one click from a structured appraisal.
Section 2
How papers are ranked, and why that's honest
Relevance is computed from free, auditable signals, not a paid per-article pass by a language model that nobody can inspect.
Provenance
Which specialty's query actually returned the article. A paper pulled by the nephrology query is, by construction, about nephrology. This is the strongest and cheapest signal there is.
Evidence tier
Taken from PubMed's own PublicationTypePubMed's own metadata label for what a paper IS, randomized trial, meta-analysis, guideline, review, and so on. metadata, not inferred. A guideline, meta-analysis or randomized trial outranks a commentary, so the bar means something.
Semantic fit
A rerankerA model that scores how well a passage answers a specific query. It is slower but sharper than embedding similarity, so it is used to refine an already-shortlisted set. (rerank-2.5)
compares what the paper is actually about against a description of the specialty, correcting the baseline
where a query was too broad.
What the number is, and what it isn't
You set the bar; only what clears it reaches your feed. The default sort is newest-first, and an optional Impact sort blends citation impact (RCRRelative Citation Ratio: a citation metric from the NIH that is normalized for field and age, so 1.0 is roughly the median NIH-funded paper. It lets you compare impact fairly across specialties and years.) with a recency decay, so strong older work can surface without burying today's papers. The critical caveat: relevance is topical centrality, not a quality grade. A perfectly ranked paper can still be a badly designed study. Judging that is what the appraisal workspace is for, and that judgment stays yours.
Section 3
Evidence signals on every card
Each badge traces to a public database. Nothing here is inferred by a language model.
| Signal | What it tells you | Source |
|---|---|---|
| Evidence type | What kind of study this is: guideline, meta-analysis, RCT, observational, preprint or commentary. | PubMed |
| Citation impact | Citation count and the field- and time-normalized RCRRelative Citation Ratio: a citation metric from the NIH that is normalized for field and age, so 1.0 is roughly the median NIH-funded paper. It lets you compare impact fairly across specialties and years., where 1.0 ≈ the median NIH-funded paper. | NIH iCiteAn NIH tool that reports citation counts and the Relative Citation Ratio for PubMed articles. → OpenAlexA free, open catalog of scholarly papers and their citations, used here as a fallback source for citation data and to tell which papers are open access anywhere. |
| Retraction safety | Retractions, expressions of concern and errata, flagged loudly. The AI librarian will never cite a retracted paper as live evidence. | PubMed correction links |
| Funding source | Who paid for the study. A fast bias signal, hidden until you ask so it never clutters the card. | Europe PMCA free European database of life-science articles and open-access full text, the primary source for funding details and full-text mining. |
| Open-access full text | Whether the methods and results, rather than only the abstract, are available to read and to ground the AI in. | Europe PMC → COREA free global aggregator of open-access research from repositories worldwide. Used as a second source of full text for open-access papers that are not in Europe PMC. |
Section 4
Exactly where AI is used, and where it can be wrong
This is the section most tools leave out. The overwhelming majority of what you see is database records and arithmetic. AI is confined to a short, nameable list, so here is that list in both directions.
✓ No AI involved
- The daily fetch, de-duplication and the feed itself
- Evidence type, citation counts and RCR, retraction flags, funding
- Keyword and PubMed search, and every filter
- The appraisal checklists and frameworks (RoB 2, Newcastle-Ottawa, QUADAS-2, AMSTAR-2, AGREE II)
- Radar's items, topic tags and importance ranking
- Saves, notes, folders, sessions, exports and sign-off
⚠ AI involved
- Relevance rating: a model's estimate of how on-topic a paper is (never a quality grade)
- Study-design label: suggested, and you confirm or correct it
- The librarian & appraisal coach: grounded in the indexed text, and they cite sources
- Appraisal drafts: a starting point from the abstract, never your finished work
- Guideline quick-reference cards and Radar desk briefings: generated text
- Compare (beta): judges which studies test the same question
How the librarian stays grounded
Corpus reads the indexed abstracts and, for open-access papers, the actual full text (fetched from Europe PMCA free European database of life-science articles and open-access full text, the primary source for funding details and full-text mining. and COREA free global aggregator of open-access research from repositories worldwide. Used as a second source of full text for open-access papers that are not in Europe PMC., so it can quote the methods and results rather than the summary), points you at its sources, and is instructed to use the paper's real figures and never invent them. It works four ways: ask about one paper, ask across your library, prep a journal club, and explain a statistic you tap. It reports what the papers say. It does not give medical advice.
Section 5
Search & discovery
One bar, two languages
Plain language searches by meaning; native PubMed syntax (field tags, AND/OR/NOT) runs as a real PubMed query. Search reaches all of PubMed, not only the indexed corpus.
Related papers
Nearest neighbours by cosine similarityA measure of how close two embeddings point in the same direction, from 0 (unrelated) to 1 (identical meaning). It is how "related papers" and duplicate detection are computed. over the stored embeddings, so any article becomes a jumping-off point. No AI call: it's arithmetic on vectors already computed.
Tracked topics
Describe a clinical question in plain English and it becomes a standing search by meaning, surfacing new matches as the library grows each day.
Quality-first by default
The library and search default to trials, meta-analyses and guidelines (the studies worth discussing), with one click to widen to everything.
Guidelines library
Practice guidelines kept out of the article feed and organized by specialty, each with an AI-formatted quick-reference card (labeled as such) so you can skim scope and recommendations before opening the source.
Landmark trials
A per-specialty library of the trials that changed practice, for orientation in a field that is new to you.
Section 6
Radar: the world around the literature
Foreground is built to consume the science; Radar covers the regulatory and industry world next to it, using the same filterable approach. It is a quiet, optional tool, not the main event.
Radar pulls FDA approvals and clearances (including De Novo first-of-kind devices), recalls with hazard severity, drug shortages, new trials, Medicare coverage decisions, society and regulator guidance, and healthtech news, all from free public sources (openFDAThe U.S. Food and Drug Administration's free public API for drug and device approvals, recalls, and shortages., ClinicalTrials.govThe U.S. public registry of clinical trials, a source for newly posted and updated studies., CMS, ex-US regulators including the EMA and MHRA, CDC MMWR, AHA/ACC). Every item is tagged by topic (authoritatively where possible, from FDA review panels, the NLM drug-class vocabulary and FDA's own AI-enabled-device list, with keyword matching as a fallback), scored by a transparent rule-based importance formula (hazard, event type, approval pathway, cross-source corroboration, recency), de-duplicated across sources, and cross-linkable to related papers already in your library. Topics include a dedicated AI in Medicine desk.
The one AI piece is the desk briefings: once a day, per desk, that desk's new items are grouped into themes with a short practice implication. It is generated once and shared by everyone (never per click), and the past week stays readable. The items, tags and ranking underneath are not AI.
Section 7
Your journal club
A club is a shared space: one feed, shared saves, group notes, and structured appraisals (PICOA standard way to frame a clinical question: Population, Intervention, Comparison, Outcome., design, bias domains, verdict) that each member drafts privately and can publish to the club when ready. Organize saves into project folders, schedule sessions around papers with a presenter and agenda, submit an appraisal to a faculty member for sign-off on a milestone scale, export a folder's citations (BibTeX / RIS / NBIBBibTeX, RIS, and NBIB are standard citation file formats that reference managers (Zotero, EndNote, Mendeley) can import directly. or formatted AMA), and share a folder as a public read-only link. Appraisal drafts stay private to their author until shared. Publishing the club copy is a deliberate act.
Section 8
The stack, and what it costs
The whole application is a single FlaskA lightweight Python web framework. The whole app runs as a single Flask process. process over SQLiteA small, fast database that lives in a single file on disk, with no separate database server to run or maintain.: small, inspectable, and cheap to run. The corpus, the FTS5SQLite's built-in full-text search engine. It indexes every word in the titles and abstracts so keyword search returns instantly. keyword index, the vectors and every club's data live in files on one disk, backed up off-site daily. There is no cluster and no separate database server.
Daily surveillance and ranking run with no language model and no API key. The AI features run on the operator's key behind hard, app-wide monthly ceilings (currently a $200/month cap): when the ceiling is reached, paid AI pauses and everything else keeps working. That is why total spend stays bounded no matter how many people use it, and why the app can be free today. As usage grows this may move to a low monthly subscription, or annual adoption by programs and departments, in exchange for expanded capability. For now it is open.
Section 9
Honest about the limits
- Relevance is topical centrality, not a quality grade. A high score says a paper is on-topic, nothing more.
- Search reaches all of PubMed, but ranking and embeddings cover the curated corpus (today 140,186 articles and growing), which favors recent clinical work.
- Most summaries draw on the abstract and the indexed record, which can lag or omit detail. Full text is only available for open-access papers.
- Retraction flags follow PubMed and can trail a real retraction by weeks.
- Citation metrics lag too. A paper published this week has no meaningful citation signal yet.
It helps you find, frame and organize. The appraisal is always yours.
Read the shorter version, or just try it
The about page is a one-screen overview, and why it works covers the education evidence behind the format.
Start a club, free See the demoBuilt on public data: PubMed and the NCBI E-utilities, NIH iCite, Europe PMC, OpenAlex, ClinicalTrials.gov, and openFDA. Open-access full text is aggregated by CORE (Knoth et al., CORE: A Global Aggregation Service for Open Access Papers, Nature Scientific Data 10, 366, 2023).